Understand the Underlying Structure and Algo

The front end.
Post Reply
Gardiola
Posts: 29
Joined: Mon Mar 11, 2013 4:38 pm

Understand the Underlying Structure and Algo

Post by Gardiola »

Good morning,

I am a new user of Genie (and soon of Smile).

So far, I am impressed by your work and software.

I would like to know the different feature of algo in order to learn the structure.

Do you have a link that shows the advantage and weakness of each methodology ?

By adance, thanks for this.

Didier
marek [BayesFusion]
Site Admin
Posts: 430
Joined: Tue Dec 11, 2007 4:24 pm

Re: Understand the Underlying Structure and Algo

Post by marek [BayesFusion] »

Didier,

I guess it's a long story and I'm afraid my answer will not satisfy you. There are two classes of general algorithms: constraint search-based (in GeNIe and SMILE, the only algorithm of this class is the PC algorithm) and Bayesian search (here we have Bayesian search, Essential Graph Search, and Greedy Thick Thinning). The two classes differ fundamentally in their approach and may generally produce different results but there are theorems showing that they are equivalent in the long run (i.e., under ideal circumstances). Constraint-based search produces a class of graphs and Bayesian search produces an instance of a class. For causal discovery I prefer constraint-based search because I know precisely what the algorithm was able to find and which arrows are impossible to orient. We have recently added a function that derives the class from a directed graph (essentially for the same purpose). You can verify the final output of your algorithm on your original data using cross-validation (use validate from the Data/Learning menu. The general game is learning a graphical representation of the joint probability distribution among the variables and this is a hard task, almost like looking for a needle in a hay stack. The remaining three algorithms learn naive Bayes models, which are an approximation of the structure of the joint p.d., often good enough in terms of the resulting classification accuracy of the model and useful especially when your data set is small (fewer parameters to learn). Does this help or confuse you further :-)?
Cheers,

Marek
Gardiola
Posts: 29
Joined: Mon Mar 11, 2013 4:38 pm

Re: Understand the Underlying Structure and Algo

Post by Gardiola »

Marek,

Thanks for your email.

I think your answer is usefull.

I tested your proposition and ideas on some cases: it allowed me to start my understanding of Genie.

Thanks for this !!!!

Didier

ps: I am still interested in the different methodology of learning and graph construction (If you have any kind of insigths like these one, feel free to let me know)
Post Reply